Bank Governors, Prime Ministers, and Teenagers: When Deepfakes Target Real People

Community Smart Hub · AI & You Series

Published 22 June 2026  ·  5 min read  ·  No tech background needed

This week, deepfake technology hit the headlines in the most personal ways possible — a bank governor investigating a fake video of himself, a head of government warning the public about a scam video using his likeness, and a shocking number of teenagers targeted by AI-generated intimate images. This blog explains what happened, why it matters, and what better technology might do about it.

1. “Real vs Fake” Detection Is Getting a Proper Workout

Researchers have built a new AI detector — called FlowFake — that can spot fake voices even when it has never heard that particular AI voice generator before. Remarkably, it achieves this with just 34,000 internal settings (called parameters).

Parameters — Think of these as the dials and knobs inside an AI model that get fine-tuned during training. The more dials, the more powerful (and expensive) the model. FlowFake has just 34,000; a modern large language model like ChatGPT has billions.

Despite its small size, FlowFake outperformed much larger systems at detecting fake speech across different recording conditions — which is important, because real-world fake audio does not always come from the same source as the training data.

Plain English: Imagine a small, nimble security guard who is actually better at spotting fraudsters than a much larger team — because they have been trained to look for the right clues.

2. How We Measure “How Good Is a Deepfake Detector?” Is Broken

Scientists published research this week showing that the standard test used to score deepfake detectors — called AUC — can be misleading. A detector can score highly on the test but still fail badly in the real world.

AUC (Area Under the Curve) — A score between 0 and 1 that measures how well a system separates “real” from “fake.” A score of 1 is perfect; 0.5 means it’s guessing randomly. The problem: a high AUC on one dataset does not mean it works on a different dataset from a different source.

Why this matters: If governments or courts rely on AI detectors that have only been tested in a lab, and those detectors fail in the wild, the consequences could be serious — from wrongful fraud decisions to false evidence in court.

The Problem with Lab-Only Testing🔬 LAB TESTDetector trained onDataset A (controlled)✅Score: 98% — Great!Passes testReal world🌍 REAL WORLDTested on Dataset B(different sources,mixed platforms)❌Fails badlyLesson:Test in mixed,real-world conditions

3. When a Watermark Does Not Mean What You Think It Means

A research team published a striking finding: it is possible to manipulate what an AI says while leaving its watermark completely intact. In tests involving AI-generated financial advice, the AI’s output was secretly skewed — but watermark-checking tools still declared the content as “verified.”

Watermark (in AI) — A hidden pattern embedded in AI-generated text or images that proves it came from a specific AI system. Like a wax seal on a letter — it shows who sent it, but not whether the letter tells the truth.

Critical warning: A watermark proves that an AI produced the content. It does not prove that the content is accurate, honest, or unmanipulated. Never treat a watermark as a guarantee of truth.

4. The EU Compliance Deadline Is Now Official

Companies that use AI to generate or manipulate content can now sign up to the EU’s voluntary Code of Practice on AI transparency. Those who sign by 22 July 2026 will be listed as early adopters. Those who do not sign must still prove they are meeting the rules independently when they come into force on 2 August.

Code of Practice — A voluntary agreement setting out what responsible companies promise to do. Think of it like a commitment to a professional standard — similar to a builder signing up to a trade association’s code of conduct.

In practice: From August, if you are a company releasing AI content and you have not signed up or proved you have proper labelling in place, you could face regulatory action.

5. Real-Life Headlines This Week

This was not just a week of technical reports. It was a week of real harm to real people:

The Bank of England governor was reported to be investigating a deepfake video purporting to show him in a confrontation with Nigel Farage. The video was fake. The investigation is real.

The Irish Taoiseach (Prime Minister) issued a public warning after a deepfake video appeared online showing him apparently endorsing a financial scam. He urged the public to “be vigilant.”

Teenagers across the UK and US reported being targeted by AI-generated intimate images — created without their knowledge or consent. Research suggests this is happening at a scale that is “disturbingly easy” to achieve with freely available tools.

Who Is Being Harmed by Deepfakes Right NowAI Deepfake Tool(freely available)Politiciansshown in fakecompromising scenesPublic officialsused in scam videosTeenagerstargeted by AIintimate imagesOrdinary peoplemisled by fake reviews

---

📌 The Bottom Line

Deepfakes are no longer just a research topic. They are being used against politicians, public figures, and — most worryingly — young people. Better technology to detect fakes is on the way, but the technology to create them is advancing faster. This makes legal protections, public awareness, and community education more important than ever.

If you or someone you know has been targeted by a deepfake, contact the Revenge Porn Helpline (UK): 0345 6000 459.Community Smart Hub · AI & You Series · Blog 3 of 5 · Week of 20 June 2026
Sources: ICML 2026, The Times, The Sun, NY Post, EU Code of Practice, new research on AUC metrics